Chap 8

Chapter 8
Stationarity, Cointegration Tests
‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 1/94

Stationarity and Unit Root Testing Why do we
need to test for Non-Stationarity?
• The stationarity or otherwise of a series can strongly influence
its behaviour and properties - e.g. persistence of shocks will
be infinite for nonstationary series
• Spurious regressions. If two variables are trending over time, a
regression of one on the other could have a high R 2 even if
the two are totally unrelated
• If the variables in the regression model are not stationary, then
it can be proved that the standard assumptions for asymptotic
analysis will not be valid. In other words, the usual “t-ratios”
will not follow a t-distribution, so we cannot validly undertake
hypothesis tests about the regression parameters.

Value of R 2 for 1000 Sets of Regressions of a
Non-stationary Variable on another Independent
Non-stationary Variable
200
160
120
frequency
80
40
0
0.00 0.25 0.50 0.75
2

Value of t-ratio on Slope Coefficient for 1000 Sets
of Regressions of a Non-stationary Variable on
another Independent Non-stationary Variable
120
100
80
frequency
60
40
20
0
–750 –500 –250 0 250 500 750
t-ratio
Two types of Non-Stationarity
• Various definitions of non-stationarity exist
• In this chapter, we are really referring to the weak form or

covariance stationarity
• There are two models which have been frequently used to
characterise non-stationarity: the random walk model with
drift:
yt = µ + yt−1 + ut (1)
and the deterministic trend process:
yt = α + βt + ut (2)
where ut is iid in both cases.

Stochastic Non-Stationarity
• Note that the model (1) could be generalised to the case
where yt is an explosive process:
yt = µ + φyt−1 + ut
where φ > 1.
• Typically, the explosive case is ignored and we use φ = 1 to
characterise the non-stationarity because
– φ > 1 does not describe many data series in economics and
finance.
– φ > 1 has an intuitively unappealing property: shocks to the
system are not only persistent through time, they are
propagated so that a given shock will have an increasingly
large influence.

Stochastic Non-stationarity: The Impact of Shocks
• To see this, consider the general case of an AR(1) with no
drift:
yt = φyt−1 + ut (3)
Let φ take any value for now.

• We can write:
yt−1 = φyt−2 + ut−1

yt−2 = φyt−3 + ut−2
• Substituting into (3) yields
yt = φ(φyt−2 + ut−1 ) + ut
= φ2 yt−2 + φut−1 + ut

Stochastic Non-stationarity: The Impact of Shocks
(Cont’d)
• Substituting again for yt−2
yt = φ2 (φyt−3 + ut−2 ) + φut−1 + ut

= φ3 yt−3 + φ2 ut−2 + φut−1 + ut
• Successive substitutions of this type lead to:
yt = φT y0 + φut−1 + φ2 ut−2 + φ3 ut−3 + · · ·

+φT u0 + ut

The Impact of Shocks for Stationary and
Non-stationary Series
• We have 3 cases:
(1) φ < 1 ⇒ φT → 0 as T → ∞
So the shocks to the system gradually die away.
(2) φ = 1 ⇒ φT = 1 ∀ T
So shocks persist in the system and never die away. We obtain
∞
X
yt = y0 + ut as T →∞
t=0
So the current value of y is just an infinite sum of past shocks

plus some starting value of y0 .
(3) φ > 1. Now given shocks become more influential as time
goes on, since if φ > 1, φ3 > φ2 > φ, etc.

Detrending a Stochastically Non-stationary Series
• Going back to our 2 characterisations of non-stationarity, the

r.w. with drift:
yt = µ + yt−1 + ut (4)
and the trend-stationary process
yt = α + βt + ut (5)
• The two will require different treatments to induce

stationarity. The second case is known as deterministic
non-stationarity and what is required is detrending.

Detrending a Stochastically Non-stationary Series
(Cont’d)
• The first case is known as stochastic non-stationarity, where

there is a stochastic trend in the data. Letting
∆yt = yt − yt−1 and Lyt = yt−1 so that
(1 − L) yt = yt − Lyt = yt − yt−1 . If (4) is taken and yt−1
subtracted from both sides
yt − yt−1 = µ + ut
∆ yt = µ + ut
We say that we have induced stationarity “by differencing

one”.

Detrending a Series: Using the Right Method
• Although trend-stationary and difference-stationary series are

both “trending” over time, the correct approach needs to be
used in each case.
• If we first difference the trend-stationary series, it would
“remove” the non-stationarity, but at the expense on
introducing an MA(1) structure into the errors.
• Conversely if we try to detrend a series which has stochastic
trend, then we will not remove the non-stationarity.
• We will now concentrate on the stochastic non-stationarity
model since deterministic non-stationarity does not adequately
describe most series in economics or finance.

Sample Plots for various Stochastic Processes: A
White Noise Process
–1 1 40 79 118 157 196 235 274 313 352 391 430 469
–2
–3
–4

Random Walk and a Random Walk with Drift
70
60
Random walk
50
Random walk with drift
40
30
20
10
0
1 19 37 55 73 91 109 127 145 163 181 199 217 235 253 271 289 307 325 343 361 379 397 415 433 451 469 487
–10
–20

Deterministic Trend Process
30
25
20
15
10
0
1 40 79 118 157 196 235 2 74 313 352 391 430 469
–5

Autoregressive Processes with differing values of φ
(0, 0.8, 1)
15
10
Phi = 1
Phi = 0.8
Phi = 0
5
0
1 53 105 157 209 261 313 365 417 469 521 573 625 677 729 784 833 885 937 989
–5
–10
–15
–20

Definition of Non-Stationarity
• Consider again the simplest stochastic trend model:
yt = yt−1 + ut
or
∆yt = ut
• We can generalise this concept to consider the case where the

series contains more than one “unit root”. That is, we would
need to apply the first difference operator, ∆, more than once
to induce stationarity.
Definition

Definition of Non-Stationarity (Cont’d)
If a non-stationary series, yt must be differenced d times

before it becomes stationary, then it is said to be integrated of
order d. We write yt ∼ I(d). So if yt ∼ I(d) then
∆d yt ∼ I(0).
An I(0) series is stationary
An I(1) series contains one unit root
yt = yt−1 + ut

Characteristics of I(0), I(1) and I(2) Series
• An I(2) series contains two unit roots and so would require

differencing twice to induce stationarity.
• I(1) and I(2) series can wander a long way from their mean
value and cross this mean value rarely.
• I(0) series should cross the mean frequently.
• The majority of economic and financial series contain a single

unit root, although some are stationary and consumer prices
have been argued to have 2 unit roots.

How do we test for a unit root?
• The early and pioneering work on testing for a unit root in
time series was done by Dickey and Fuller (Dickey and Fuller
1979, Fuller 1976). The basic objective of the test is to test
the null hypothesis that φ =1 in:
yt = φyt−1 + ut
against the one-sided alternative φ < 1. So we have

H0 : series contains a unit root
versus H1 : series is stationary.
• We usually use the regression:
∆yt = ψyt−1 + ut
so that a test of φ = 1 is equivalent to a test of ψ = 0 (since

φ − 1 = ψ).
Different forms for the DF Test Regressions
• Dickey Fuller tests are also known as τ tests: τ , τµ , ττ .
• The null (H0 ) and alternative (H1 ) models in each case are
i. H0 : yt = yt−1 + ut
H1 : yt = φyt−1 + ut , φ<1
This is a test for a random walk against a stationary
autoregressive process of order one (AR(1))
ii. H0 : yt = yt−1 + ut
H1 : yt = φyt−1 + µ + ut , φ<1
This is a test for a random walk against a stationary AR(1)
with drift.
iii. H0 : yt = yt−1 + ut
H1 : yt = φyt−1 + µ + λt + ut , φ<1
This is a test for a random walk against a stationary AR(1)
with drift and a time trend.
Computing the DF Test Statistic
• We can write
∆yt = ut
where ∆yt = yt − yt−1 , and the alternatives may be expressed

as
∆yt = ψyt−1 + µ + λt + ut
with µ = λ = 0 in case i), and λ = 0 in case ii) and

ψ = φ − 1. In each case, the tests are based on the t-ratio on
the yt−1 term in the estimated regression of ∆yt on yt−1 , plus
a constant in case ii) and a constant and trend in case iii).
The test statistics are defined as

Computing the DF Test Statistic (Cont’d)
ψ̂
test statistic =
SE ˆ( ˆψ)
• The test statistic does not follow the usual t-distribution
under the null, since the null is one of non-stationarity, but
rather follows a non-standard distribution. Critical values are
derived from Monte Carlo experiments in, for example, Fuller
(1976). Relevant examples of the distribution are shown in
table 4.1 below

Critical Values for the DF Test
Significance level 10% 5% 1%

CV for constant but no trend −2.57 −2.86 −3.43
CV for constant and trend −3.12 −3.41 −3.96
The null hypothesis of a unit root is rejected in favour of the

stationary alternative in each case if the test statistic is more
negative than the critical value.

The Augmented Dickey Fuller (ADF) Test
• The tests above are only valid if ut is white noise. In
particular, ut will be autocorrelated if there was
autocorrelation in the dependent variable of the regression
(∆yt ) which we have not modelled. The solution is to
“augment” the test using p lags of the dependent variable.
The alternative model in case (i) is now written:
p
X
∆yt = ψyt−1 + αi ∆yt−i + ut
i=1
• The same critical values from the DF tables are used as

before. A problem now arises in determining the optimal
number of lags of the dependent variable
• There are 2 ways
– use the frequency of the data to decide
– use information criteria
Testing for Higher Orders of Integration
• Consider the simple regression:
∆yt = ψyt−1 + ut
We test that H0 : ψ = 0 vs. H1 : ψ < 0.

• If H0 is rejected, we simply conclude that yt does not contain
a unit root.
• But what do we conclude if H0 is not rejected? The series
contains a unit root, but is that it? No! What if yt ∼ I(2)?
We would still not have rejected. So we now need to test
H0 : yt ∼ I(2) vs. H1 : yt ∼ I(1)
We would continue to test for a further unit root until we

rejected H0 .
Testing for Higher Orders of Integration (Cont’d)
• We now regress ∆2 yt on ∆yt−1 (plus lags of ∆2 yt if

necessary).
• Now we test H0 : ∆yt ∼ I (1) which is equivalent to H0 :

yt ∼ I (2).
• So in this case, if we do not reject (unlikely), we conclude

that yt is at least I(2).

The Phillips-Perron Test
• Phillips and Perron have developed a more comprehensive

theory of unit root nonstationarity. The tests are similar to
ADF tests, but they incorporate an automatic correction to
the DF procedure to allow for autocorrelated residuals.
• The tests usually give the same conclusions as the ADF tests,
and the calculation of the test statistics is complex.

Criticism of Dickey-Fuller and Phillips-Perron-type
tests
• Main criticism is that the power of the tests is low if the
process is stationary but with a root close to the
non-stationary boundary.
e.g. the tests are poor at deciding if φ=1 or φ=0.95,
especially with small sample sizes.
• If the true data generating process (dgp) is
yt = 0.95yt−1 + ut
then the null hypothesis of a unit root should be rejected.

• One way to get around this is to use a stationarity test as well
as the unit root tests we have looked at.

Stationarity tests
• Stationarity tests have
H0 : yt is stationary
versus H1 : yt is non-stationary
So that by default under the null the data will appear
stationary.
• One such stationarity test is the KPSS test (Kwaitowski,
Phillips, Schmidt and Shin, 1992).
• Thus we can compare the results of these tests with the
ADF/PP procedure to see if we obtain the same conclusion.
ADF / PP KPSS
• A Comparison H0 : yt ∼ I (1) H0 : yt ∼ I (0)
H1 : yt ∼ I (0) H0 : yt ∼ I (1)

Stationarity tests (Cont’d)
• 4 possible outcomes
Reject H0 and Do not reject H0
Do not Reject H0 and Reject H0
Reject H0 and Reject H0
Do not reject H0 and Do not reject H0

Unit Root Tests with Structural Breaks
• The standard Dickey-Fuller-type unit root tests presented

above do not perform well if there are structural breaks in the
series
• The tests have low power in such circumstances and they fail
to reject the unit root null hypothesis when it is incorrect as
the slope parameter in the regression of yt on yt−1 is biased
towards unity
• The larger the break and the smaller the sample, the lower the
power of the test
• Unit root tests are also oversized in the presence of structural
breaks

Unit Root Tests with Structural Breaks (Cont’d)
• Perron (1989) demonstrates that after allowing for structural

breaks in the tests, a whole raft of macroeconomic series may
be stationary
• He argues that most economic time series are best

characterised by broken trend stationary processes, i.e. a
deterministic trend but with a structural break.

The Perron (1989) Procedure - Background
• Perron (1989) proposes three test equations differing
dependent on the type of break that is thought to be present:
1. A ‘crash’ model that allows a break in the level (i.e. the
intercept)
2. A ‘changing growth’ model that allows for a break in the
growth rate (i.e. the slope)
3. A model that allows for both types of break to occur at the
same time, changing both the intercept and the slope of the
trend.
• Define the break point in the data as Tb and Dt is a dummy

variable defined as
(
0 if t < Tb
Dt =
1 if t ≥ Tb

The Perron (1989) Procedure - Details
• The equation for the third (most general) version of the test is
p
X
∆yt = ψyt−1 + µ + α1 Dt + α2 (t − Tb )Dt + λt + αi ∆yt−i + ut
i=1
• For the crash only model, set α2 = 0

• For the changing growth only model, set α1 = 0
• In all three cases, there is a unit root with a structural break
at Tb under the null hypothesis and a series that is a
stationary process with a break under the alternative
• A limitation of this approach is that it assumes that the break
date is known in advance
• It is possible, however, that the date will not be known and
must be determined from the data.
The Banerjee et al. (1992) and Zivot and Andrews
(1992) Procedures - Background
• More seriously, Christiano (1992) has argued that the critical
values employed with the test will presume the break date to
be chosen exogenously
• But most researchers will select a break point based on an
examination of the data
• Thus the asymptotic theory assumed will no longer hold
• Banerjee et al. (1992) and Zivot and Andrews (1992)

introduce an approach to testing for unit roots in the presence
of structural change that allows the break date to be selected
endogenously.
• Their methods are based on recursive, rolling and sequential
tests.
(1992) Procedures - Details
• For the recursive and rolling tests, Banerjee et al. propose
four specifications.
1. The standard Dickey-Fuller test on the whole sample
2. The ADF test conducted repeatedly on the sub-samples and
the minimal DF statistic is obtained
3. The maximal DF statistic obtained from the sub-samples
4. The difference between the maximal and minimal statistics
• For the sequential test, the whole sample is used each time
with the following regression being run
p
X
∆yt = ψyt−1 + µ + ατt (tused ) + λt + αi ∆yt−i + ut
i=1
where tused = Tb /T .
(1992) Procedures – Details 2
• The test is run repeatedly for different values of Tb over as

much of the data as possible (a ‘trimmed sample’)
• This excludes the first few and the last few observations
• Clearly it is τt (tused ) allows for the break, which can either be

in the level (where τt (tused ) = 1 if t > tused and 0 otherwise);
or the break can be in the deterministic trend (where
τt (tused ) = t − tused if t > tused and 0 otherwise)
• For each specification, a different set of critical values is
required, and these can be found in Banerjee et al.

Further Extensions
• Perron (1997) proposes an extension of the Perron (1989)

technique but using a sequential procedure that estimates the
test statistic allowing for a break at any point during the
sample to be determined by the data
• This technique is very similar to that of Zivot and Andrews,
except that his is more flexible since it allows for a break
under both the null and alternative hypotheses
• A further extension would be to allow for more than one
structural break in the series – for example, Lumsdaine and
Papell (1997) enhance the Zivot and Andrews (1992)
approach to allow for two structural breaks.

Testing for Unit Roots with Structural Breaks
Example: EuroSterling Interest Rates
• Brooks and Rew (2002) examine whether EuroSterling
interest rates are best viewed as unit root process or not,
allowing for the possibility of structural breaks in the series
• Failure to account for structural breaks (caused, for example,
by changes in monetary policy or the removal of exchange
rate controls) may lead to incorrect inferences regarding the
validity or otherwise of the expectations hypothesis.
• Their sample covers the period 1 January 1981 to 1
September 1997
• They use the standard Dickey-Fuller test, the recursive and
sequential tests of Banerjee et al. They also employ the rolling
test, the Perron (1997) approach and several other techniques
Testing for Unit Roots with Structural Breaks in
EuroSterling Interest Rates – Results
Maturity Recursive statistics Sequential statistics

tDF max
t̂DF min
t̂DF diff
t̂DF min
t̃DF min
t̃DF
,trend ,mean
Short rate -2.44 -1.33 -3.29 1.96 -2.99 -4.79
7-days -1.95 -1.33 -3.19 1.86 -2.44 -5.65
1-month -1.82 -1.07 -2.90 1.83 -2.32 -4.78
3-months -1.80 -1.02 -2.75 1.73 -2.28 -4.02
6-months -1.86 -1.00 -2.85 1.85 -2.28 -4.10
1-year -1.97 -0.74 -2.88 2.14 -2.35 -4.55
Critical values -3.13 -1.66 -3.88 3.21 -4.11 -4.58
Notes: Source: Brooks and Rew (2002), taken from tables 1, 4 and 5. t̃DF min
,trend denotes the sequential test
min
statistic allowing for a break in the trend, while t̃DF ,mean is the test statistic allowing for a break in the level.
The final row presents the 10% level critical values for each type of test obtained from Banerjee et al.
(1992, p. 278, table 2).

EuroSterling Interest Rates – Conclusions
• The findings for the recursive tests are that the unit root null
should not be rejected at the 10% level for any of the
maturities examined
• For the sequential tests, the results are slightly more mixed
with the break in trend model not rejecting the null
hypothesis, while it is rejected for the short, 7-day and the
1-month rates when a structural break is allowed for in the
mean

EuroSterling Interest Rates – Conclusions (Cont’d)
• The weight of evidence indicates that short term interest rates

are best viewed as unit root processes that have a structural
break in their level around the time of ‘Black Wednesday’ (16
September 1992) when the UK dropped out of the European
Exchange Rate Mechanism
• The longer term rates, on the other hand, are I(1) processes
with no breaks

Seasonal Unit Roots
• It is possible that a series may contain seasonal unit roots, so
that it requires seasonal differencing to induce stationarity
• We would use the notation I(d,D) to denote a series that is
integrated of order d,D and requires differencing d times and
seasonal differencing D times to obtain a stationary process
• Osborn (1990) develops a test for seasonal unit roots based
on a natural extension of the Dickey-Fuller approach.
• However, Osborn also shows that only a small proportion of
macroeconomic series exhibit seasonal unit roots; the majority
have seasonal patterns that can better be characterised using
dummy variables, which may explain why the concept of
seasonal unit roots has not been widely adopted

Cointegration: An Introduction
• In most cases, if we combine two variables which are I(1),

then the combination will also be I(1)
• More generally, if we combine variables with differing orders of
integration, the combination will have an order of integration
equal to the largest. i.e.,
if Xi,t ∼ I(di ) for i = 1, 2, 3, . . . , k
so we have k variables each integrated of order di .
Let
k
X
zt = αi Xi,t (6)
i=1
Then zt ∼ I(max di ).

Linear Combinations of Non-stationary Variables
• Rearranging (6), we can write
k
X
X1,t = βi Xi,t + zt0
i=2
where βi = − αα1i , zt0 = zt

α1 , i = 2, . . . , k.
• This is just a regression equation.
• But the disturbances would have some very undesirable

properties: zt0 is not stationary and is autocorrelated if all of
the Xi are I(1).
• We want to ensure that the disturbances are I(0). Under what
circumstances will this be the case?

Definition of Cointegration (Engle & Granger,
1987)
• Let zt be a k × 1 vector of variables, then the components of
zt are cointegrated of order (d,b) if
i. All components of zt are I(d)
ii. There is at least one vector of coefficients α such that
α0 zt ∼ I(d − b)
• Many time series are non-stationary but “move together” over
time.
• If variables are cointegrated, it means that a linear
combination of them will be stationary.
• There may be up to r linearly independent cointegrating
relationships (where r ≤ k − 1), also known as cointegrating
vectors. r is also known as the cointegrating rank of zt .
• A cointegrating relationship may also be seen as a long term
relationship.
Cointegration and Equilibrium
• Examples of possible Cointegrating Relationships in finance:

– spot and futures prices
– ratio of relative prices and an exchange rate
– equity prices and dividends
• Market forces arising from no arbitrage conditions should

ensure an equilibrium relationship.
• No cointegration implies that series could wander apart
without bound in the long run.

Equilibrium Correction or Error Correction Models
• When the concept of non-stationarity was first considered, a
usual response was to independently take the first differences
of a series of I(1) variables.
• The problem with this approach is that pure first difference
models have no long run solution.
e.g. Consider yt and xt both I(1).
The model we may want to estimate is
∆yt = β∆xt + ut
But this collapses to nothing in the long run.
• The definition of the long run that we use is where
yt = yt−1 = y ; xt = xt−1 = x.
• Hence all the difference terms will be zero, i.e.
∆yt = 0; ∆xt = 0.
Specifying an ECM
• One way to get around this problem is to use both first
difference and levels terms, e.g.
∆yt = β1 ∆xt + β2 (yt−1 − γxt−1 ) + ut (7)
• yt−1 − γxt−1 is known as the error correction term.
• Provided that yt and xt are cointegrated with cointegrating

coefficient γ, then (yt−1 − γxt−1 ) will be I(0) even though the
constituents are I(1).
• We can thus validly use OLS on (7)
• The Granger representation theorem shows that any

cointegrating relationship can be expressed as an equilibrium
correction model.

Testing for Cointegration in Regression
• The model for the equilibrium correction term can be
generalised to include more than two variables:
yt = β1 + β2 x2t + β3 x3t + · · · + βk xkt + ut (8)
• ut should be I(0) if the variables yt , x2t , . . . xkt are

cointegrated
• So what we want to test is the residuals of equation (8) to see
if they are non-stationary or stationary. We can use the
DF/ADF test on ut .
So we have the regression
∆ût = ψ ût−1 + vt with vt ∼ iid.
• However, since this is a test on the residuals of an actual

model, ût , then the critical values are changed.
Testing for Cointegration in Regression:
Conclusions
• Engle and Granger (1987) have tabulated a new set of critical

values and hence the test is known as the Engle Granger
(E.G.) test.
• We can also use the Durbin Watson test statistic or the
Phillips Perron approach to test for non-stationarity of ût .
• What are the null and alternative hypotheses for a test on the
residuals of a potentially cointegrating regression?
H0 : unit root in cointegrating regression’s residuals
H1 : residuals from cointegrating regression are stationary

Methods of Parameter Estimation in Cointegrated
Systems: The Engle-Granger Approach
• There are (at least) 3 methods we could use: Engle Granger,
Engle and Yoo, and Johansen.
• The Engle Granger 2 Step Method
This is a single equation technique which is conducted as
follows:
Step 1:
– Make sure that all the individual variables are I(1).
– Then estimate the cointegrating regression using OLS.
– Save the residuals of the cointegrating regression, .
– Test these residuals to ensure that they are I(0).
Step 2:

Methods of Parameter Estimation in Cointegrated
Systems: The Engle-Granger Approach (Cont’d)
– Use the step 1 residuals as one variable in the error correction

model e.g.
∆yt = β1 ∆xt + β2 (ût−1 ) + ut
where ût−1 = yt−1 − τ̂ xt−1 .

An Example of a Model for Non-stationary
Variables: Lead-Lag Relationships between Spot and
Futures Prices
Background
• We expect changes in the spot price of a financial asset and its
corresponding futures price to be perfectly contemporaneously
correlated and not to be cross-autocorrelated.
corr(∆log(ft ), ∆ ln(st )) ≈ 1
corr(∆log(ft ), ∆ ln(st−k )) ≈ 0 ∀ k > 0
corr(∆log(ft−j ), ∆ ln(st )) ≈ 0 ∀j >0
• We can test this idea by modelling the lead-lag relationship
between the two.
• We will consider two papers Tse(1995) and Brooks et al
(2001).
Futures & Spot Data
• Tse (1995): 1055 daily observations on NSA stock index and

stock index futures values from December 1988 - April 1993.
• Brooks et al (2001): 13,035 10-minutely observations on the
FTSE 100 stock index and stock index futures prices for all
trading days in the period June 1996 – 1997.

Methodology
• The fair futures price is given by
∗ (r −d)(T −t)
Ft = St e
where Ft∗ is the fair futures price, St is the spot price, r is a

continuously compounded risk-free rate of interest, d is the
continuously compounded yield in terms of dividends derived
from the stock index until the futures contract matures, and
(T − t) is the time to maturity of the futures contract.
Taking logarithms of both sides of equation above gives
∗
ft = st +(r − d)(T − t)
• First, test ft and st for nonstationarity.

Dickey-Fuller Tests on Log-Prices and Returns for
High Frequency FTSE Data
Futures Spot
Dickey–Fuller statistics −0.1329 −0.7335
for log-price data
Dickey–Fuller statistics −84.9968 −114.1803
for returns data

Cointegration Test Regression and Test on
Residuals
• Conclusion: log Ft and log St are not stationary, but ∆ log Ft
and ∆ log St are stationary.
• But a model containing only first differences has no long run
relationship.
• Solution is to see if there exists a cointegrating relationship
between ft and st which would mean that we can validly
include levels terms in this framework.
• Potential cointegrating regression:
st = γ0 + γ1 ft + zt
where zt is a disturbance term.
• Estimate the regression, collect the residuals, Zˆt , and test
whether they are stationary.
Estimated Equation and Test for Cointegration for
High Frequency FTSE Data
Coefficient Estimated value

γ̂0 0.1345
γ̂1 0.9834
DF test on residuals Test statistic
ẑt −14.7303
Source: Brooks, Rew and Ritson (2001).

Conclusions from Unit Root and Cointegration
Tests
• Conclusion: Zˆt are stationary and therefore we have a

cointegrating relationship between log Ft and log St .
• Final stage in Engle-Granger 2-step method is to use the first
stage residuals, Zˆt as the equilibrium correction term in the
general equation.
• The overall model is
∆ log st = β0 + δ ẑt−1 + β1 ∆ ln st−1 + α1 ∆ ln ft−1 + vt

Estimated Error Correction Model for High
Frequency FTSE Data
Coefficient Estimated value t-ratio

β̂ 0 9.6713E−06 1.6083
δ̂ −0.8388 −5.1298
β̂ 1 0.1799 19.2886
α̂1 0.1312 20.4946
Look at the signs and significances of the coefficients:

• α̂1 is positive and highly significant
• β̂ 1 is positive and highly significant
• δ̂ is negative and highly significant

Forecasting High Frequency FTSE Returns
• Is it possible to use the error correction model to produce

superior forecasts to other models?
Comparison of Out of Sample Forecasting Accuracy

ECM ECM-COC ARIMA VAR
RMSE 0.0004382 0.0004350 0.0004531 0.0004510
MAE 0.4259 0.4255 0.4382 0.4378
% Correct direction 67.69% 68.75% 64.36% 66.80%

Can Profitable Trading Rules be Derived from the
ECM-COC Forecasts?
• The trading strategy involves analysing the forecast for the
spot return, and incorporating the decision dictated by the
trading rules described below. It is assumed that the original
investment is £1000, and if the holding in the stock index is
zero, the investment earns the risk free rate.
– Liquid Trading Strategy - making a round trip trade (i.e. a
purchase and sale of the FTSE100 stocks) every ten minutes
that the return is predicted to be positive by the model.
– Buy-&-Hold while Forecast Positive Strategy - allows the
trader to continue holding the index if the return at the next
predicted investment period is positive.
– Filter Strategy: Better Predicted Return Than Average -
involves purchasing the index only if the predicted returns are
greater than the average positive return.

Can Profitable Trading Rules be Derived from the
ECM-COC Forecasts? (Cont’d)
– Filter Strategy: Better Predicted Return Than First Decile -

only the returns predicted to be in the top 10% of all returns
are traded on
– Filter Strategy: High Arbitrary Cut Off - An arbitrary filter of
0.0075% is imposed,

Spot Trading Strategy Results for Error Correction
Model Incorporating the Cost of Carry
Terminal Terminal Return(%)

Wealth Return(%) Wealth (£) Annualised Number
Trading strategy (£) annualised with slippage with slippage of trades
Passive investment 1040.92 4.09 1040.92 4.09 1
{49.08} {49.08}
Liquid trading 1156.21 15.62 1056.38 5.64 583
{187.44} {67.68}
Buy-and-Hold while 1156.21 15.62 1055.77 5.58 383
forecast positive {187.44} {66.96}
Filter I 1144.51 14.45 1123.57 12.36 135
{173.40} {148.32}
Filter II 1100.01 10.00 1046.17 4.62 65
{120.00} {55.44}
Filter III 1019.82 1.98 1003.23 0.32 8
{23.76} {3.84}

Conclusions
• The futures market “leads” the spot market because:

• the stock index is not a single entity, so
• some components of the index are infrequently traded
• it is more expensive to transact in the spot market
• stock market indices are only recalculated every minute
• Spot & futures markets do indeed have a long run
relationship.
• Since it appears impossible to profit from lead/lag
relationships, their existence is entirely consistent with the
absence of arbitrage opportunities and in accordance with
modern definitions of the efficient markets hypothesis.

The Engle-Granger Approach: Some Drawbacks
This method suffers from a number of problems:
1. Unit root and cointegration tests have low power in finite
samples
2. We are forced to treat the variables asymmetrically and to
specify one as the dependent and the other as independent
variables.
3. Cannot perform any hypothesis tests about the actual
cointegrating relationship estimated at stage 1.
– Problem 1 is a small sample problem that should disappear
asymptotically.
– Problem 2 is addressed by the Johansen approach.
– Problem 3 is addressed by the Engle and Yoo approach or the
Johansen approach.
The Engle & Yoo 3-Step Method
• One of the problems with the EG 2-step method is that we

cannot make any inferences about the actual cointegrating
regression.
• The Engle & Yoo (EY) 3-step procedure takes its first two
steps from EG.
• EY add a third step giving updated estimates of the

cointegrating vector and its standard errors.
• The most important problem with both these techniques is

that in the general case above, where we have more than two
variables which may be cointegrated, there could be more
than one cointegrating relationship.

The Engle & Yoo 3-Step Method (Cont’d)
In fact there can be up to r linearly independent cointegrating
vectors (where r ≤ g − 1), where g is the number of variables
in total.
• So, in the case where we just had y and x, then r can only be
one or zero.
• But in the general case there could be more cointegrating
relationships.
• And if there are others, how do we know how many there are
or whether we have found the “best”?
• The answer to this is to use a systems approach to
cointegration which will allow determination of all r
cointegrating relationships - Johansen’s method.

Testing for and Estimating Cointegrating Systems
Using the Johansen Technique Based on VARs
• To use Johansen’s method, we need to turn the VAR of the
form
yt = β1 yt−1 + β2 yt−2 +··· + βk yt−k +

g ×1 g ×g g ×1 g ×g g ×1 g ×g g ×1
into a VECM, which can be written as
∆yt = Πyt−k + Γ1 ∆yt−1 + Γ2 ∆yt−2 + · · · + Γk−1 ∆yt−(k−1)

+ut
where Π = ( ki=1 βi ) − Ig and Γi = ( ij=1 βj ) − Ig

P P
Π is a long run coefficient matrix since all the ∆yt−i = 0.

Review of Matrix Algebra necessary for the
Johansen Test
• Let Π denote a g × g square matrix and let c denote a g×1
non-zero vector, and let λ denote a set of scalars.
• λ is called a characteristic root or set of roots of Π if we can
write
Πc = λc
g ×g g ×1 g ×1
• We can also write

Πc = λIg c
and hence
(Π − λIg )c = 0

Johansen Test (Cont’d)
where Ig is an identity matrix, and hence
• Since c 6= 0 by definition, then for this system to have zero

solution, we require the matrix (Π − λI g ) to be singular (i.e.
to have zero determinant).
|Π − λI g | = 0
• For example, let Π be the 2 × 2 matrix

5 1
Π=
2 4

• Then the characteristic equation is

5 1 1 0
|Π − λI p | = −λ =0
2 4 0 1

5−λ 1
= = (5 − λ)(4 − λ) − 2 = λ2 − 9λ + 18
2 4−λ
• This gives the solutions λ = 6 and λ = 3.
• The characteristic roots are also known as Eigenvalues.
• The rank of a matrix is equal to the number of linearly

independent rows or columns in the matrix.

• We write Rank(Π) = r
• The rank of a matrix is equal to the order of the largest

square matrix we can obtain from Π which has a non-zero
determinant.
• For example, the determinant of Π above 6= 0, therefore it has

rank 2.

The Johansen Test and Eigenvalues
• Some properties of the eigenvalues of any square matrix A:
1. the sum of the eigenvalues is the trace
2. the product of the eigenvalues is the determinant
3. the number of non-zero eigenvalues is the rank
• Returning to Johansen’s test, the VECM representation of the

VAR was
∆yt = Πyt−1 + Γ1 ∆yt−1 + Γ2 ∆yt−2 + · · · + Γk−1 ∆yt−(k−1)

+ut
• The test for cointegration between the y ’s is calculated by

looking at the rank of the Π matrix via its eigenvalues (to
prove this requires some technical intermediate steps).

The Johansen Test and Eigenvalues (Cont’d)
• The rank of a matrix is equal to the number of its
characteristic roots (eigenvalues) that are different from zero.
• The eigenvalues denoted λi are put in order:
λ1 ≥ λ2 ≥ λg
• If the variables are not cointegrated, the rank of Π will not be
significantly different from zero, so λi = 0 ∀ i.
Then if λi = 0, ln(1 − λi ) = 0
if the λ ’s are roots, they must be less than 1 in absolute
value.
• Say rank (Π) = 1, then ln(1 − λ1 ) will be negative and
ln(1 − λi ) = 0
• If the eigenvalue λi is non-zero, then
ln(1 − λi ) < 0 ∀ i > 1.
The Johansen Test Statistics
• The test statistics for cointegration are formulated as
g
X
λtrace (r ) = −T ln(1 − λ̂i )
i=r +1
and
λmax (r , r + 1) = −T ln(1 − λ̂r +1 )
where λ̂ is the estimated value for the ith ordered eigenvalue

from the Π matrix.
λtrace tests the null that the number of cointegrating vectors
is less than equal to r against an unspecified alternative.
λtrace = 0 when all the λi = 0, so it is a joint test.
λmax tests the null that the number of cointegrating vectors is
r against an alternative of r + 1.
Decomposition of the Matrix
• For any 1 < r < g , Π is defined as the product of two

matrices:
Π = αβ 0
g ×g g ×r r ×g
• β contains the cointegrating vectors while α gives the

“loadings” of each cointegrating vector in each equation.

Decomposition of the Matrix (Cont’d)
• For example, if g = 4 and r = 1, α and β will be 4 × 1, and
Πyt−k will be given by:
   
α11 y1
 α12   y2 
Π=  α13  ( β11 β12 β13 β14 ) 
  
y3 
α14 y4 t−k
or


α11
 α12 
 α13  ( β11 y1 + β12 y2 + β13 y3 + β14 y4 )t−k
Π= 
α14

Johansen Critical Values
• Johansen & Juselius (1990) provide critical values for the 2

statistics. The distribution of the test statistics is
non-standard. The critical values depend on:
1. the value of g − r , the number of non-stationary components
2. whether a constant and/or trend are included in the
regressions.
• If the test statistic is greater than the critical value from

Johansen’s tables, reject the null hypothesis that there are r
cointegrating vectors in favour of the alternative that there
are more than r.

The Johansen Testing Sequence
• The testing sequence under the null is r = 0, 1, . . . , g − 1
• so that the hypotheses for λtrace are
H0 : r = 0 versus H1 : 0 < r ≤ g
H0 : r = 1 versus H1 : 1 < r ≤ g
H0 : r = 2 versus H1 : 2 < r ≤ g
.. .. ..
. . .
H0 : r = g − 1 versus H1 : r = g
• We keep increasing the value of r until we no longer reject the

null.

Interpretation of Johansen Test Results
• But how does this correspond to a test of the rank of the Π

matrix?
• r is the rank of Π.
• Π cannot be of full rank (g) since this would correspond to

the original yt being stationary.
• If Π has zero rank, then by analogy to the univariate case,
∆yt depends only on ∆yt−j and not on yt−1 , so that there is
no long run relationship between the elements of yt−1 . Hence
there is no cointegration.
• For 1 < rank(Π) < g , there are multiple cointegrating
vectors.

Hypothesis Testing Using Johansen
• EG did not allow us to do hypothesis tests on the

cointegrating relationship itself, but the Johansen approach
does.
• If there exist r cointegrating vectors, only these linear

combinations will be stationary.
• You can test a hypothesis about one or more coefficients in

the cointegrating relationship by viewing the hypothesis as a
restriction on the Π matrix.
• All linear combinations of the cointegrating vectors are also

cointegrating vectors.

Hypothesis Testing Using Johansen (Cont’d)
• If the number of cointegrating vectors is large, and the

hypothesis under consideration is simple, it may be possible to
recombine the cointegrating vectors to satisfy the restrictions
exactly.
• As the restrictions become more complex or more numerous,

it will eventually become impossible to satisfy them by
renormalisation.
• After this point, if the restriction is not severe, then the

cointegrating vectors will not change much upon imposing the
restriction.

Hypothesis Testing Using Johansen (Cont’d)
• A test statistic to test this hypothesis is given by

r
X
test statistic = −T [ln(1 − λi ) − ln(1 − λi ∗ )] ∼ χ2 (m)
i=1
where
λ∗i are the characteristic roots of the restricted model,
λi are the characteristic roots of the unrestricted model,
r is the number of non-zero characteristic roots in the
unrestricted model and m is the number of restrictions.

Cointegration Tests using Johansen: Three
Examples
Example 1: Hamilton(1994, pp.647 )

• Does the PPP relationship hold for the US / Italian exchange
rate - price system?
• A VAR was estimated with 12 lags on 189 observations. The
Johansen test statistics were
r λmax critical value
0 22.12 20.8
1 10.19 14.0
• Conclusion: there is one cointegrating relationship.

Example 2: Purchasing Power Parity (PPP)
• PPP states that the equilibrium exchange rate between 2

countries is equal to the ratio of relative prices
• A necessary and sufficient condition for PPP is that the log of
the exchange rate between countries A and B, and the logs of
the price levels in countries A and B be cointegrated with
cointegrating vector [1 –1 1].
• Chen (1995) uses monthly data for April 1973-December 1990
to test the PPP hypothesis using the Johansen approach.

Cointegration Tests of PPP with European Data
Tests for
cointegration between r =0 r ≤ 1 r ≤ 2 α1 α2
FRF–DEM 34.63∗ 17.10 6.26 1.33 −2.50
FRF–ITL 52.69∗ 15.81 5.43 2.65 −2.52
FRF–NLG 68.10∗ 16.37 6.42 0.58 −0.80
FRF–BEF 52.54∗ 26.09∗ 3.63 0.78 −1.15
DEM–ITL 42.59∗ 20.76∗ 4.79 5.80 −2.25
DEM–NLG 50.25∗ 17.79 3.28 0.12 −0.25
DEM–BEF 69.13∗ 27.13∗ 4.52 0.87 −0.52
ITL–NLG 37.51∗ 14.22 5.05 0.55 −0.71
ITL–BEF 69.24∗ 32.16∗ 7.15 0.73 −1.28
NLG–BEF 64.52∗ 21.97∗ 3.88 1.69 −2.17
Critical values 31.52 17.95 8.18 – –
Notes: FRF – French franc; DEM – German mark; NLG – Dutch guilder; ITL – Italian lira; BEF – Belgian franc.
Source: Chen (1995). Reprinted with the permission of Taylor & Francis Ltd <www.tandf.co.uk>.

Example 3: Are International Bond Markets
Cointegrated?
• Mills & Mills (1991)
• If financial markets are cointegrated, this implies that they

have a “common stochastic trend”.
• Data:
Daily closing observations on redemption yields on government
bonds for 4 bond markets: US, UK, West Germany, Japan.
• For cointegration, a necessary but not sufficient condition is
that the yields are nonstationary. All 4 yields series are I(1).

Testing for Cointegration Between the Yields
• The Johansen procedure is used. There can be at most 3
linearly independent cointegrating vectors.
• Mills & Mills use the trace test statistic:
g
X
λtrace (r ) = −T ln(1 − λ̂i )
i=r +1
where λi are the ordered eigenvalues.

Johansen Tests for Cointegration between International Bond Yields
Critical values
r (number of cointegrating
vectors under the null hypothesis) Test statistic 10% 5%
0 22.06 35.6 38.6
1 10.58 21.2 23.8
2 2.52 10.3 12.0
3 0.12 2.9 4.2
Source: Mills and Mills (1991). Reprinted with the permission of Blackwell Publishers.

Testing for Cointegration Between the Yields
(Cont’d)
• Conclusion: No cointegrating vectors.
• The paper then goes on to estimate a VAR for the first
differences of the yields, which is of the form
k
X
∆Xt = Γi ∆Xt−i + vt
i=1
where:
     
X (US)t Γ11i Γ12i Γ13i Γ14i v1t
 X (UK)t  Γ21i Γ22i Γ23i Γ24i   v2t 
Xt = 
 X (WG)t  , Γi = Γ31i
,v = 
Γ34i  t  v3t
  
Γ32i Γ33i 
X (JAP)t Γ41i Γ42i Γ43i Γ44i v4t
They set k = 8
Variance Decompositions for VAR of International
Bond Yields
Explained by movements in
Explaining Days
movements in ahead US UK Germany Japan
US 1 95.6 2.4 1.7 0.3
5 94.2 2.8 2.3 0.7
10 92.9 3.1 2.9 1.1
20 92.8 3.2 2.9 1.1
UK 1 0.0 98.3 0.0 1.7
5 1.7 96.2 0.2 1.9
10 2.2 94.6 0.9 2.3
20 2.2 94.6 0.9 2.3
Germany 1 0.0 3.4 94.6 2.0
5 6.6 6.6 84.8 3.0
10 8.3 6.5 82.9 3.6
20 8.4 6.5 82.7 3.7
Japan 1 0.0 0.0 1.4 100.0
5 1.3 1.4 1.1 96.2
10 1.5 2.1 1.8 94.6
20 1.6 2.2 1.9 94.2
Impulse Responses for VAR of International Bond
Yields
Response of US to innovations in
Days after shock US UK Germany Japan
0 0.98 0.00 0.00 0.00
1 0.06 0.01 −0.10 0.05
2 −0.02 0.02 −0.14 0.07
3 0.09 −0.04 0.09 0.08
4 −0.02 −0.03 0.02 0.09
10 −0.03 −0.01 −0.02 −0.01
20 0.00 0.00 −0.10 −0.01
Response of UK to innovations in
0 0.19 0.97 0.00 0.00
1 0.16 0.07 0.01 −0.06
2 −0.01 −0.01 −0.05 0.09
3 0.06 0.04 0.06 0.05
4 0.05 −0.01 0.02 0.07
10 0.01 0.01 −0.04 −0.01
20 0.00 0.00 −0.01 0.00
Response of Germany to innovations in
0 0.07 0.06 0.95 0.00
1 0.13 0.05 0.11 0.02
2 0.04 0.03 0.00 0.00
3 0.02 0.00 0.00 0.01
4 0.01 0.00 0.00 0.09
10 0.01 0.01 −0.01 0.02
20 0.00 0.00 0.00 0.00
Response of Japan to innovations in
0 0.03 0.05 0.12 0.97
1 0.06 0.02 0.07 0.04
2 0.02 0.02 0.00 0.21
3 0.01 0.02 0.06 0.07
4 0.02 0.03 0.07 0.06
10 0.01 0.01 0.01 0.04
20 0.00 0.00 0.00 0.01

Chap 8

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chap 8

Uploaded by

Copyright:

Available Formats

Chapter 8

Stationarity, Cointegration Tests

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 1/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 2/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 3/94

• In this chapter, we are really referring to the weak form or

and the deterministic trend process:

where ut is iid in both cases.

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 5/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 6/94

Let φ take any value for now.

yt−1 = φyt−2 + ut−1

• Substituting into (3) yields

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 7/94

• Substituting again for yt−2

yt = φ2 (φyt−3 + ut−2 ) + φut−1 + ut

• Successive substitutions of this type lead to:

yt = φT y0 + φut−1 + φ2 ut−2 + φ3 ut−3 + · · ·

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 8/94

So the current value of y is just an infinite sum of past shocks

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 9/94

• Going back to our 2 characterisations of non-stationarity, the

and the trend-stationary process

• The two will require different treatments to induce

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 10/94

• The first case is known as stochastic non-stationarity, where

We say that we have induced stationarity “by differencing

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 11/94

• Although trend-stationary and difference-stationary series are

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 12/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 13/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 14/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 15/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 16/94

• Consider again the simplest stochastic trend model:

• We can generalise this concept to consider the case where the

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 17/94

If a non-stationary series, yt must be differenced d times

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 18/94

• An I(2) series contains two unit roots and so would require

• The majority of economic and financial series contain a single

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 19/94

against the one-sided alternative φ < 1. So we have

• We usually use the regression:

so that a test of φ = 1 is equivalent to a test of ψ = 0 (since

where ∆yt = yt − yt−1 , and the alternatives may be expressed

with µ = λ = 0 in case i), and λ = 0 in case ii) and

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 22/94

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 23/94

Significance level 10% 5% 1%

The null hypothesis of a unit root is rejected in favour of the

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 24/94

• The same critical values from the DF tables are used as

We test that H0 : ψ = 0 vs. H1 : ψ < 0.

H0 : yt ∼ I(2) vs. H1 : yt ∼ I(1)

We would continue to test for a further unit root until we

• We now regress ∆2 yt on ∆yt−1 (plus lags of ∆2 yt if

• Now we test H0 : ∆yt ∼ I (1) which is equivalent to H0 :

• So in this case, if we do not reject (unlikely), we conclude

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 27/94

• Phillips and Perron have developed a more comprehensive

‘Introductory Econometrics for Finance’ by Dr. Le Trung Thanh 2020 28/94